Optimal Res-UNET architecture with deep supervision for tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor segmentation is crucial for diagnosis and treatment planning in medical imaging.
• Deep learning (DL) models, especially U-Net, show promise but face optimization challenges.
• Enhancing U-Net variants for better performance and computational efficiency is an ongoing research area.

Purpose of the Study:

• To develop an optimized Residual U-Net (Res-UNET) architecture incorporating deep supervision.
• To improve the accuracy of brain tumor segmentation in MRI datasets.
• To address limitations of conventional segmentation methods.

Main Methods:

• Evaluated multiple U-Net variations, including Res-UNET and attention-enhanced U-Net.
• Utilized the BraTS 2018 public MRI dataset for training and validation.
• Integrated deep supervision and employed Dice loss with focal loss for data imbalance.
• Conducted ablation studies to analyze encoder complexity, filter count, and post-processing.

Main Results:

• The proposed Res-UNET with deep supervision achieved a high average Dice score of 0.9498 via cross-validation.
• Post-processing enhanced segmentation robustness, particularly for small tumor regions.
• Res-UNET demonstrated superior accuracy and faster training times compared to standard U-Net.

Conclusions:

• Optimized Res-UNET with deep supervision significantly boosts brain tumor MRI segmentation accuracy.
• The model effectively handles dataset imbalance and computational inefficiencies.
• Further research should explore these optimized U-Net variants in other medical imaging segmentation tasks.